November 20, 2021
The machine learning company OpenAI now provides access to its public API without a waiting list. This means that anyone interested can register for the interface to the GPT-3 language model, provided they come from the right countries. The programming AI Codex, which is based on GPT-3, is also publicly available via the API.
From its own point of view, OpenAI has now implemented sufficient security measures to prevent misuse of the language model. GPT-3 is considered an extremely powerful speech generator, but like all machine learning models, it draws its knowledge from human input.
In this way, the model not only adopts people’s grammatical knowledge in its texts, but also their prejudices. Overall, this bias is a problem in machine learning applications. Back in 2018, Amazon had crushed its ML-based evaluation system for applications because it had discriminated against female applicants compared to male applicants.
An analysis of GPT-3 earlier this year concluded that the model harbored deeply ingrained prejudices against Muslims. This was taken from the innumerable texts on the Internet that were used to train the model.
The security measures include a content filter that currently still has beta status and texts as
unsafe classifies. In addition, see the documentation best practices for safe handling with the API.
After all, OpenAI Guidelines for using the API published, which in a certain way specify a traffic light system – from applications that are almost always approved, to evaluations that are dependent on the individual case, to prohibited applications. The latter include applications that create automated tweets or Instagram posts or otherwise share content on social media. The use of the API to influence political decisions is also taboo.
Guidelines were already in place during the non-public phase. In September, a developer had to switch off his GPT-3 chatbot because he had violated the rules of OpenAI and, among other things, had not integrated a filter or monitoring technology against abusive topics.
In principle, the API is now open to all developers, provided they do not come from the wrong countries. Instead of a tab list, OpenAI has a Liste der “Supported Countries” released. It was to be expected that North Korea, Iran, Afghanistan and China would be absent from the list, and the absence of Russia is hardly surprising either. Ukraine is also not on the positive list, but Iraq and the Holy See are explicitly listed separately.
Next to GPT-3, OpenAI opens according to the overview page for the API also the interface to the Codex model, which translates natural language into code. It’s the foundation for GitHub Copilot and started private beta in August. on the Codex project page however, the waiting list is still linked.
OpenAI was founded in 2015 as a research project, Elon Musk was one of the founders. Initially, the organization published many reinforcement learning projects. The language model GPT-3, which stands as an acronym for Generative Pre-trained Transformer 3, brought OpenAI widespread attention. In 2019, OpenAI said goodbye to the pure non-profit business and is now investing itself in start-ups in the field of machine learning – most recently with 100 million US dollars.
Meanwhile, other companies are developing alternatives to GPT-3. Microsoft and Nvidia introduced the Megatron-Turing Natural Language Generation Model (MT-NLG) in October, and the German company Aleph Alpha is relying on a multimodal model that is considered the European answer to GPT-3. Meanwhile, GPT-4 is already in progress, which, like its predecessor, will probably be a pure text model.
Further details on opening the API can be found from the OpenAI blog. If you want to use GPT-3, you first have to create an account. OpenAI has a number provided by sample applications, which should serve as starting aid.
Published in Technology
November 20, 2021